Re: [tcpm] Thank you for the QUIC session in tcpm

[Jana Iyengar <[e-mail]>]

Thanks for the clarification, Neal!

On Thu, Nov 15, 2018 at 9:15 AM Neal Cardwell <[e-mail]> wrote:

> On Wed, Nov 14, 2018 at 11:18 PM Jana Iyengar <[e-mail]> wrote:
>
>> Thanks for the clarification, Eric. I'm a bit surprised to see srtt be
>> used by both Cubic and BBR, but that discussion's not for this thread.
>>
>
> Just to add context (sorry to dilute the thread!), the spot in Linux TCP
> BBR that Eric pointed to that is using srtt is the initialization code
> (bbr_init_pacing_rate_from_rtt()). The code is using srtt here to compute a
> pacing rate (via pacing_rate = gain * cwd / srtt) largely because Linux TCP
> can hot-swap congestion control modules. Because of that possibility, the
> cwnd and srtt that BBR inherits at initialization may have been from a
> buffer-bloated CUBIC/Reno connection, where srtt may have bloated in
> proportion to cwnd, so that cwd / srtt is generally a better approximation
> to the available bandwidth than cwnd / min_rtt.  As you would imagine,
> after that initialization, Linux TCP BBR does not use srtt (or min_rtt) to
> update the pacing rate; it uses a function of the estimated bandwidth.
>
> best,
> neal
>
>
>
>> On Wed, Nov 14, 2018 at 9:44 PM Eric Dumazet <[e-mail]>
>> wrote:
>>
>>>
>>>
>>> On 11/12/2018 11:13 PM, Jana Iyengar wrote:
>>> > On what to do with an unconfirmed RTO: So, my thought is that if
>>> you've hit an RTO, it's been a REALLY LONG time, because we've gone through
>>> TLP(s) by now and failed. (This might change if we decide to combine TLP
>>> and RTO, but that's for a later discussion.)  It's not fully clear to me
>>> what the right responses here are, but some thoughts follow.
>>> >
>>> > 1/ If this was an actual increase in the RTT, it would naturally roll
>>> into the SRTT and RTTVAR. We might even consider resetting the RTT
>>> estimator, but we already know that the EWMA estimator we love takes just a
>>> while to catch on, and I don't want to litigate changing the estimator
>>> fundamentally here.
>>> >
>>> > 2/ Linux fq uses min_rtt IIRC, which shouldn't change as a result of
>>> an increase in RTT. This *might* be a good time to kick the min_rtt
>>> estimator to pace more conservatively.
>>>
>>> linux fq does not use min_rtt, but the pacing rate set by TCP.
>>>
>>> pacing rate for Cubic is using SRTT (
>>> https://elixir.bootlin.com/linux/latest/source/net/ipv4/tcp_input.c#L806
>>> )
>>>
>>> For BBR, pacing rate depends on various factors and phases, but SRTT is
>>> used, not min_rtt
>>>
>>> https://elixir.bootlin.com/linux/latest/source/net/ipv4/tcp_bbr.c#L231
>>>
>>>
>>> >
>>> > 3/ I'm not convinced that it makes sense to change the cwnd. It seems
>>> clear that an RTT increase with the same BW should increase the cwnd if
>>> anything, so it seems reasonable to continue with the same cwnd at least.
>>> If the BW changed, hopefully that'll be reflected in a congestion signal,
>>> but in the absence of a congestion signal, I think we should keep the cwnd
>>> as is.
>>> >
>>> >
>>> > On Tue, Nov 13, 2018 at 12:35 AM Praveen Balasubramanian <
>>> [e-mail] <mailto:[e-mail]>> wrote:
>>> >
>>> >     Yuchung this might be worth experimenting with and getting more
>>> data on. Maybe there should be some reaction to an unconfirmed RTO even if
>>> it is not reduction all the way to LW? I like Jana's suggestion that the
>>> implementation SHOULD drop the cwnd even for unconfirmed RTO if not pacing.
>>> >
>>> >     I checked the latest draft and pacing is a RECOMMENDED. I am
>>> noticing that transport algorithms in Linux and in QUIC are evolving
>>> assuming pacing which is a function usually implemented and configured
>>> outside of the transport.. So the transport is potentially making an
>>> assumption which may result in safety issues. For example BBR congestion
>>> control may not work very well if pacing was disabled or misconfigured? I
>>> don’t know of any robust solution to this other than build pacing into the
>>> transport itself. Pacing is also very challenging for low latency flows
>>> because of lack of support for fine grained timers.
>>> >
>>> >     -----Original Message-----
>>> >     From: Yuchung Cheng <[e-mail] <mailto:[e-mail]>>
>>> >     Sent: Monday, November 12, 2018 8:37 AM
>>> >     To: Jana Iyengar <[e-mail] <mailto:[e-mail]>>
>>> >     Cc: Praveen Balasubramanian <[e-mail] <mailto:
>>> [e-mail]>>; Bob Briscoe <[e-mail] <mailto:
>>> [e-mail]>>; Ian Swett <[e-mail] <mailto:
>>> [e-mail]>>; [e-mail] <mailto:[e-mail]> Extensions <
>>> [e-mail] <mailto:[e-mail]>>
>>> >     Subject: Re: [tcpm] Thank you for the QUIC session in tcpm
>>> >
>>> >     On Mon, Nov 12, 2018 at 4:13 AM, Jana Iyengar <[e-mail]
>>> <mailto:[e-mail]>> wrote:
>>> >     > Praveen,
>>> >     >
>>> >     > The point you're raising -- that we've lost the ack clock after
>>> an RTO
>>> >     > -- is a reasonable point. My argument is that with pacing, cwnd
>>> >     > reduction is unwarranted because in the extreme case, this
>>> collapses
>>> >     > to restart after idle, where sending at cwnd with pacing is
>>> reasonable
>>> >     I agree w/ Jana as well - the most generic way to treat burst due
>>> to prior_inflight << cwnd is pacing. With QUIC's approach, when RTO fires,
>>> cwnd remains unchanged until the ACK of the first retransmission (i.e. a
>>> probe packet) comes back. Therefore the delay is one RTT and the potential
>>> damage is an additional cwnd-worth of burst.
>>> >
>>> >     Yes the worst case is more aggressive, and it can be too much for
>>> DC incast case if pacing isn't supported - one idea is to selective enable
>>> that if RTT variance is very large vs RTT.
>>> >
>>> >     >
>>> >     > The draft does not say that a sender should reduce the cwnd if
>>> it is
>>> >     > not pacing, which we should add. Does that make sense?
>>> >     >
>>> >     > - jana
>>> >     >
>>> >     > On Fri, Nov 9, 2018 at 8:34 AM Praveen Balasubramanian
>>> >     > <pravb=[e-mail] <mailto:
>>> [e-mail]>> wrote:
>>> >     >>
>>> >     >> Yuchung I brought that difference up in an email to the quic wg
>>> >     >> earlier this week.
>>> >     >>
>>> >     >> In app send limited case, inflight could be very small compared
>>> to cwnd.
>>> >     >> So in QUIC there is potential to send a burst out after a long
>>> idle
>>> >     >> period (with outstanding data) where TCP wouldn't. The draft
>>> claims
>>> >     >> this is okay to do because RTO may have been a result of RTT
>>> increase
>>> >     >> instead of loss. Is there data to suggest on which side we
>>> should
>>> >     >> err? i.e. data on what are the chances that an RTO is due to an
>>> RTT increase versus loss.
>>> >     >>
>>> >     >> Do you see any safety concerns with delayed reduction of cwnd
>>> in case
>>> >     >> where the RTO is not spurious?
>>> >     >>
>>> >     >> -----Original Message-----
>>> >     >> From: tcpm <[e-mail] <mailto:[e-mail]>>
>>> On Behalf Of Yuchung Cheng
>>> >     >> Sent: Thursday, November 8, 2018 4:38 PM
>>> >     >> To: Bob Briscoe <[e-mail] <mailto:
>>> [e-mail]>>
>>> >     >> Cc: Ian Swett <[e-mail] <mailto:[e-mail]>>;
>>> tcpm IETF list <[e-mail] <mailto:[e-mail]>>
>>> >     >> Subject: Re: [tcpm] Thank you for the QUIC session in tcpm
>>> >     >>
>>> >     >> On Thu, Nov 8, 2018 at 3:14 AM, Bob Briscoe <
>>> [e-mail] <mailto:[e-mail]>> wrote:
>>> >     >> >
>>> >     >> > I just wanted to thank Jana for explaining QUIC loss recovery
>>> to us
>>> >     >> > (and QUIC CC as far as it goes).
>>> >     >> > And thank you Jana, Ian, the chairs of both WGs (and anyone
>>> else
>>> >     >> > involved) for setting it up.
>>> >     >> >
>>> >     >> > If one is not full-time on QUIC, it's very difficult to keep
>>> up
>>> >     >> > with all the changes. But now we have a checkpoint to start
>>> from, I
>>> >     >> > feel I will not be wasting people's time if I try to get
>>> involved -
>>> >     >> > at least I only might say something un-QUIC occasionally,
>>> rather
>>> >     >> > than nearly always. This has allowed people who understand
>>> how TCP
>>> >     >> > cold be improved to help with QUIC, when working on QUIC
>>> isn't their day job.
>>> >     >> >
>>> >     >> > Again, Thank you.
>>> >     >> I like particularly that QUIC only reduces cwnd to 1 after the
>>> loss
>>> >     >> is confirmed not upon RTO fires. It should be very feasible for
>>> TCP
>>> >     >> (at least
>>> >     >> Linux) w/ TCP timestamps. It'll save a lot of spurious cwnd
>>> reductions!
>>> >     >>
>>> >     >> Also IMHO TCP w/ quality timestamps are almost as good as QUIC
>>> pkt-ids.
>>> >     >> Google internally uses usec. We wish we could upstream it but
>>> RFC
>>> >     >> needs to be updated.
>>> >     >>
>>> >     >> >
>>> >     >> >
>>> >     >> > Bob
>>> >     >> >
>>> >     >> >
>>> >     >> > --
>>> >     >> >
>>> ________________________________________________________________
>>> >     >> > Bob Briscoe
>>> >     >> >
>>> https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fbob
>>> >     >> > briscoe.net <http://briscoe.net>%2F&amp;data=02%7C01%7Cpravb%
>>> 40microsoft.com <http://40microsoft.com>%7C3c4d22866
>>> >     >> >
>>> 8444691c16808d648bd2ce9%7C72f988bf86f141af91ab2d7cd011db47%7C1%7C0%
>>> >     >> >
>>> 7C636776374678321324&amp;sdata=T%2BWTEc42VC%2Bz%2BsmMFyjlm37hmAwfee
>>> >     >> > buPcuMYlVhgDY%3D&amp;reserved=0
>>> >     >> >
>>> >     >> > _______________________________________________
>>> >     >> > tcpm mailing list
>>> >     >> > [e-mail] <mailto:[e-mail]>
>>> >     >> >
>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fww
>>> >     >> > w.i
>>> >     >> > etf.org <http://etf.org
>>> >%2Fmailman%2Flistinfo%2Ftcpm&amp;data=02%7C01%7Cpravb%40micr
>>> >     >> > oso
>>> >     >> > ft.com <http://ft.com
>>> >%7Cf0911eeb74d7446f424508d645dbb779%7C72f988bf86f141af91ab2d7
>>> >     >> > cd0
>>> >     >> >
>>> 11db47%7C1%7C0%7C636773207316711149&amp;sdata=K667a3IQG4rarQ%2FOfAl
>>> >     >> > yhK
>>> >     >> > QQ05Cea421rgb64DlEMvs%3D&amp;reserved=0
>>> >     >>
>>> >     >> _______________________________________________
>>> >     >> tcpm mailing list
>>> >     >> [e-mail] <mailto:[e-mail]>
>>> >     >>
>>> >     >>
>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.
>>> >     >> ietf.org <http://ietf.org
>>> >%2Fmailman%2Flistinfo%2Ftcpm&amp;data=02%7C01%7Cpravb%40micro
>>> >     >> soft.com <http://soft.com
>>> >%7C3c4d228668444691c16808d648bd2ce9%7C72f988bf86f141af91ab2d7
>>> >     >>
>>> cd011db47%7C1%7C0%7C636776374678321324&amp;sdata=33WxKh5c6qL4ln5jerpx
>>> >     >> Rhytfrj1iTK284pEqv5fWKY%3D&amp;reserved=0
>>> >     >>
>>> >     >> _______________________________________________
>>> >     >> tcpm mailing list
>>> >     >> [e-mail] <mailto:[e-mail]>
>>> >     >>
>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fwww.
>>> >     >> ietf.org <http://ietf.org
>>> >%2Fmailman%2Flistinfo%2Ftcpm&amp;data=02%7C01%7Cpravb%40micro
>>> >     >> soft.com <http://soft.com
>>> >%7C3c4d228668444691c16808d648bd2ce9%7C72f988bf86f141af91ab2d7
>>> >     >>
>>> cd011db47%7C1%7C0%7C636776374678321324&amp;sdata=33WxKh5c6qL4ln5jerpx
>>> >     >> Rhytfrj1iTK284pEqv5fWKY%3D&amp;reserved=0
>>> >
>>> >
>>> >
>>> > _______________________________________________
>>> > tcpm mailing list
>>> > [e-mail]
>>> > https://www.ietf.org/mailman/listinfo/tcpm
>>> >
>>>
>>> _______________________________________________
>>> tcpm mailing list
>>> [e-mail]
>>> https://www.ietf.org/mailman/listinfo/tcpm
>>>
>> _______________________________________________
>> tcpm mailing list
>> [e-mail]
>> https://www.ietf.org/mailman/listinfo/tcpm
>>
>